►
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
Kubernetes is pushing out of the data center into stores, networks and factories. DevOps teams are excited about Kubernetes because it provides predictable operations and a cloud-like provisioning experience on just about any infrastructure. For the last year we’ve been working with organizations looking for a lightweight implementation of Kubernetes that is easy to operate and can run on x86 and ARM infrastructure without using more than 512 MB of memory. To meet these requirements we’ve developed a new distribution of Kubernetes that is designed for teams that need to deploy applications quickly and reliably to resource-constrained environments.
In this webinar Rancher co-founders Shannon Williams and Darren Shepherd will discuss:
- The specific requirements for running Kubernetes on the edge.
- Developing an operational model for single-node and clustered edge environments
- Which parts of the Kubernetes code base we were able to remove
- Why we introduced SQLlite as an alternative to etcd
- Developing support for x86, arm64 and armhf
- Management options for hundreds of micro clusters
A
Cool
so
I'm
gonna
go
ahead
and
get
started.
Everybody
all
right,
so
we're
gonna
go
ahead
and
get
this
thing
going.
So
welcome
to
today's
CN
CF
webinar
and
we're
gonna
be
introducing
a
lightweight
Cabernets
distribution
built
for
the
edge
I'm
at
Baldwin,
director
of
cloud
native
engineering
and
ciencia
ambassador
and
I'll
be
moderating.
Today's
webinar
I'd
like
to
thank
everyone
who
is
joining
us
today.
We
would
like
to
welcome
our
presenters,
Darren
Shepherd
chief
architect,
Rancher
Labs
and
Shannon
Williams
VP
Marketing
Rancher
Labs.
If
you
guys
wanted
wave
at
everybody,
that's
awesome!
A
So
just
a
few
housekeeping
items
before
we
get
started
during
the
webinar
you're,
not
able
to
talk
and
as
an
NT
as
an
attendee.
So
we
ask
you
to
if
you
have
a
question,
please
drop
that
question
into
the
flight
box,
which
is
located
at
the
bottom
of
your
screen.
Awesome,
please
feel
free
to
drop
your
your
questions
there
and
throughout
the
the
presentation
so
that
we
can
stack
those
up
and
then,
at
the
end,
we'll
we'll
go
through
those
and
ask
those
to
the
attendees.
Sorry
to
the
presenters.
A
B
Matt
and
thanks
to
everyone
who's
here
today
looks
like
we
have
a
great
great
group,
a
great
crowd,
so
I
thought
just
to
to
warm
us
up
a
little
bit.
Darren
and
I
actually
have
been
hosting
online
meetups
and
trainings
like
this
for
the
last
four
years,
or
so,
as
part
of
you
know,
starting
rancher
and
doing
this,
so
you
know
we
we
both
have.
You
know
both
really
excited
about
today's
presentation,
because
it's
a
it's
introducing
a
new
project.
The
timing
for
this
really
worked
out
nicely.
B
We
were
we've
been
working
on
this
and
we
realized
that
we
were
supposed
to
be
presenting
on
this
Meetup
the
same
day
that
that
we
were
announcing
a
new,
a
new
district
community.
So
we
thought
what
better
thing
to
do
than
to
talk
a
little
bit
about
that
and
share
it
with
everyone
here
today,
so
I'm,
Shannon,
Williams
I'm,
one
of
the
founders
of
rancher
and
I,
run
all
of
our
customer
facing
activities,
sales
and
marketing
and
the
likes
Darren
is
my
co-founder.
B
He
is
the
chief
architect
and
a
really
lead
engineer
on
everything
we
built
over
at
Rancher.
For
those
of
you
who
don't
know
who
Rancher
is
where
where's
open
source
software
company,
we
basically
build
open-source
projects
and
and
commercialize
those
our
most
famous
one,
is
called
ranch,
which
is
a
management
platform
for
containers
and
kubernetes.
B
But
we
build
all
sorts
of
different
things
like
nitro,
ass
and
Longhorn
and
other
types
of
projects,
and
we
just
open
source
everything.
We
build
and
pay
our
bills
by
providing
support
services
for
companies
that
put
that
stuff
in
production.
So
we've
been
members
of
the
CN
CF
since
it
started
I'm
on
the
board
of
the
CN
CF
the
governing
board,
and
you
know
we
we
really
so
much
of
what
we
do
is
based
around.
B
You
know
belief
that
this
space
is
growing
very
rapidly
and
the
need
for
you
know,
continuous
innovation
around
it
is,
is
not
slowing
down
in
any
way.
So
what
we're
gonna
talk
about
today
is
has
really
kind
of
proof
of
that.
It's
how
containers
and
kubernetes
are
changing
what
we
think
of
when
we
think
of
edge
computing
and
I'll
talk
a
little
bit
to
start
us
off,
but
just
kind
of
walking
through
some
of
the
some
of
the
things
we're
seeing
in
the
market.
You
know
20,000
teams
use
ranchers.
B
Being
you
know
very
appealing,
for
you
know,
centralized
application
management
well,
then,
and
spend
most
of
today
talking
about
caitria,
so
that
a
new
micro
distribution
of
kubernetes
that
we've
developed
we'll
talk
a
little
bit
about
how
we,
how
we
were
able
to
build
it.
What
we
did
to
kind
of
cut
the
size
down
and
make
it
ideal
for
running
on
really
small
footprint
environments,
we'll
talk
about
some
changes
we
made
to
the
back
end.
You
know
by
introducing
a
sequel
like
it
is
an
optional
datastore.
B
Instead
of
instead
of
using
at
CD
exclusively,
then
we
Darrin
will
demo
quite
a
bit
a
little
bit
about
some
of
the
other
stuff
we're
working
on
around
managing
these
clusters.
Now
that
you,
once
you
get
to
the
point
where
you're
running
you
know
micro,
cluster
single
node
clusters,
how
do
you
manage
you
know
tens
of
thousands
of
these?
So
that's
kind
of
our
hope
for
today,
please,
fire
away
with
your
questions.
B
This
talk
to
themselves,
and
so
you
know,
I
pull
a
few
articles
and
places
that
you
can
go
and
and
listen
to
companies
share
their
experience,
taking
kubernetes
and
pushing
it
to
the
edge
one
of
the
one
of
the
interesting
ones
that
we
spend
a
lot
of
time.
Working
with
and
talking
to
is
the
team
of
chick-fil-a
they
they've
been
really
public
about
their
desire
to.
You
know,
basically
build
and
deploy
a
micro,
kubernetes
implementation
into
every
store,
every
chick-fil-a
location,
and
it's
a
it's
a
really
interesting
project.
B
It
gets
to
the
heart
of
why
we're
seeing
this
because
they
have
in
a
more
and
more
complex
software,
stack
that
they
want
to
run
in
in
the
store
you
know
from
analytics
to
the
actual
point-of-sale
systems
to
different
types
of
marketing
applications
to
employee
management
applications
right
there.
They
want
to
implement
more
and
more
software,
that's
more
and
more
complex
and
includes
things
like
analytics
and
monitoring,
etc
down
into
the
store-
and
you
know
historically,
this
was
all
just
servers
right.
B
They
would
have
a
server
closet
or
they
would
have
some
type
of
a
local
server
that
they
would
run
in
the
store
and
now
with
containers
and
kubernetes,
you
know
they're
their
excitement.
What
you
hear
them
talk
about
is
you
know
the
ability
to
really
know
that
there
were
applications
that
they're
deploying
are
exactly.
You
know
the
same
code
that
they
tested
in
the
lab
that
they
tested
in
you
know
in
other
deployments,
and
so
this.
A
B
A
good
example
someone
who's
been
really
public
about.
The
idea
of
you
know
bare
metal
clusters
of
kubernetes
in
the
store,
but
they're
not
they're,
not
alone,
I
mean
the
use.
Cases
are
everywhere
target.
You
know,
get
did
a
really
nice
presentation
blog
last
year,
also
in
the
summer
talking
about
their
movement
to
putting
kubernetes
in
the
edge
how
they're,
using
CI
to
push.
B
B
We
see
ton
of
this
a
good
example
of
more
of
an
industrial
automation
to
the
type
of
use
cases
one
that
was
presented
at
coop
con
and
Shanghai
and
gold
went
there,
the
second
largest
builder
and
operator
of
wind
turbines
in
the
world,
and
they
run
now
a
massive
stack
of
software
on
kubernetes
on
the
edge
everywhere
they
run.
They
run
these
wind
farms
and
I.
Think
it
gives
you
an
idea.
They
shared
their
stack
and
I.
B
Think
if
you
look
at
their
stack
now,
we've
been
working
with
this
team
for
a
long
time
as
they've
built
it
all
on
rancher,
and
we,
you
know
one
of
the
things
we
do.
That
was
always
really
clear
was
that
it
wasn't
the
fact
that
they
needed
to
get
software
out
onto
devices
that
would
cause
them
to
look
at
containers.
It
was
the
fact
that
you
know
their
applications
running
on
the
edge,
we're
starting
to
look
more
and
more
like
the
applications
they
wanted
to
run.
B
In
the
data
center,
you
can
see
I
mean
they're
running
Hadoop,
they're
running
Kafka,
running
log
stash
their
own.
You
know
they're
running
all
sorts
of
different
analytics
and
business
applications
that
are
helping
them.
You
know
optimize
the
the
actual
wind
turbines
and
optimize
for
the
energy
production
that
they're
trying
to
get
and
if
they're
now
running
I
want
to
say
2,000
or
more
clusters
around
the
world
everywhere.
B
The
running
is
wind
turbines
and
kubernetes
is
powering
all
of
these,
and
so
we're
seeing
this
just
continuous
use
cases
for
more
complex
applications,
distributed
more
frequently
out
to
more
places-
and
you
know
just
like
in
the
data
center
containers
and
kubernetes-
are
just
an
awesome
choice
for
deploying
this
type
of
complex
software
onto
especially
an
unknown
set
of
infrastructure
or
a
wide
variety
of
infrastructure,
and
you
know
it
all
comes
down
to
the
fact.
Containers
are
pretty
awesome
right.
B
It
still
gives
us
this
great
packaging,
for
you
know,
approach
for
taking
our
software
testing
it
you're,
making
it
repeatable.
You
know
pushing
it
out
to
really
you
know
the
same
reason:
it's
valuable
to
be
able
to
deploy
to
you
know,
on
premise
and
in
the
cloud
it's
valuable
to
deploy
to
lots
of
different
edge
locations.
We
found
that
you
know
there
was
when
we
talked
to
people
about
why
kubernetes
for
these
edge,
it's
all
about
the
operational
benefits
it
gives
them
right.
B
The
ability
to
restart
services,
to
you
know,
set
health
checks
to
you
know,
define
what
a
chain
looks
like
for
a
cluster
to
both
run
as
a
single
node.
In
many
cases,
it's
just
a
single
edge
device,
so
they're
running
kubernetes,
potentially
on
one
arm
server,
but
also
you
know
the
ability
to
cluster
that
if
they
need
to
run
three
servers
or
five
servers
or
seven,
they
can
kind
of
expand
or
contract
the
amount
of
infrastructure
and
have
the
same
platform
experience.
B
That's
really
kind
of
one
of
the
big
benefits
of
kubernetes
at
the
edge,
and
then
you
know
the
more
obvious
ones.
The
fact
that
kubernetes
has
a
really
rich
infrastructure,
a
set
of
infrastructure
plugins.
So
they
can,
you
know,
depending
on
what
they
need.
Is
you
saw
in
the
goldwyn
one
they
were
using
the
reason
cluster
F
asked
for
storage.
You
know.
In
other
cases,
people
are
using
the
local.
B
Using
different
types
of
storage
services,
because
often
what
they're
doing
on
the
edge
does
generate
persistent
data
that
needs
to
be
managed,
but
it's
not
just
storage.
It's
also.
You
know
the
the
fact
that
kubernetes,
it
has
good
support
from
arm,
and
so
you
can
use
it
with
different
chipsets
and
different
CPUs.
The
other
nice
thing
about
criminalities
is
the
richness
of
the
ecosystem
developing
around
it
right
from
the
monitoring
to
the
logging
to
management
platforms.
You
can
you
know,
kubernetes
is
a
standard.
B
You
can
get
a
certified
implementation
of
kubernetes
and
lots
of
cases,
and
so
it
you
know
I
can
test
on
kubernetes
in
the
cloud
I
can
run
on
kubernetes
and
on
arm
servers
running
in
places
like
Amazon.
There's,
really
cool
companies
out
there
building
building
you
know
new
arm
devices.
We've
been
working
with
this
company
called
hide
cell.
That
has
just
a
really
great
ability
to
kind
of
stack
these.
These
small
arm
servers
and
so
they're,
just
a
strong
ecosystem
developing
around
the
edge
and
a
strong
ecosystem
developing
around
kubernetes
and
containers.
B
That
makes
this
a
really
appealing
model
for
teams
that
are
trying
to
figure
out
how
to
push
code
into
ATM
machines
or
into
factories
or
into
even
satellites.
So
we
see
a
lot
lots
of
these
different
types
of
use
cases
the
the
journey
we've
been
on.
You
know,
we've
sort
of
I
think
the
first
time
we
ran
into
this
use
case.
B
You
know
people
were
trying
to
use
Rancher
itself
to
deploy,
deploy,
containers
and
kubernetes
out
to
the
edge,
and,
if
you
think
about
you
know
from
us,
you
know
before
before
we
ever
got
into
you
know
kind
of
the
idea
of
an
edge
optimized,
Esper
Nettie's
we've
been
provisioning
through
Nettie's
clusters,
but
rancher
for
a
long
time
and
so
rancher.
As
always,
we've
done
a
lot
of
this
type
of
thing,
where
users
use
Rancher
to
run
lots
of
clusters
and
lots
of
places.
B
B
You
know
we
were
also
supporting
along
hosted
kubernetes.
We
were,
you
know,
creating
it,
allowing
people
to
import
existing
certified
clusters
from
different
types
of
distros
that
they
were
setting
up,
and
as
part
of
that,
we
made
the
decision
to
break
out
something
we
call
rke,
which
was
our
our
the
kubernetes
installer
functionality
that
had
been
sort
of
bundled
into
into
ranch,
or
we
kind
of
broke
it
out
as
its
own
standalone
open-source
project
that
people
could
use
and
what
made
it
you
what's
made
it.
B
A
really
popular
project
is
that
it
has
this
sort
of
ability
to
you
know
very
easily,
define
a
cluster.
Can
big
set
up
what
you
want
to
deploy
and
how
you
want
to
deploy
it
and
then,
and
then
just
deploy
these
things
and
one
of
the
things
that
the
Daren
and
the
team
did,
that
that
kind
of
I
think
made
Arkady
very
popular
one
of
the
reasons
we
started
working
with
the
guys
chick-fil-a
is
they
were
using
our
KC
and
was
we
learned
why
one
of
the
things
that
was
was
obvious?
B
Was
they
were
using
it
because
it
allowed
them
to
launch
a
whole
bunch
of
services
and
attach
a
bunch
of
add-ons
as
part
of
employment?
And
so
you
know
they
could
not
just
sort
of
launch
kubernetes
but
also
have
kubernetes
come
up.
You
know
instantiate
itself,
and
then
you
know
deploy
a
whole
bunch
of
applications
directly
onto
that,
but
even
doing
all
that
we,
you
know,
we
kind
of
found.
There
were
challenges
and
the
challenges
were
pretty
consistent.
You
know
we're
as
we
talk
to
people,
they
were.
B
You
know
most
the
time
and
you
know
most
communities
distributions,
including
rke,
really
didn't,
have
first-class
support
for
our
more.
If
it
did,
it
was
sort
of
coming
out
months
later,
and
this
was
more
and
more
important
for
us
as
we
as
we
worked
with
people
on
on
the
on
the
edge
was
clearly
arm
was
gonna,
be
a
big
part
of
that
infrastructure.
B
Also,
it
was
just
fat.
You
know
when,
as
with
you
know,
for
any
kind
of
deployment
of
kubernetes,
we
were
often
seeing
you
know
that
it
was
eating
up
half
of
the
device.
This
was
you
know,
especially
for
a
single
node
type
of
implementation.
If
you're
consuming
four
gigs
of
ram
you're
you're
really
spending
a
lot
of
the
capacity
that
you
have
out
there
on
the
edge
on
the
management
layer-
and
you
know,
at
its
core
human
disease-
wasn't
really
ever
built
for
this
kind
of
offline
embedded,
distributed
management
right.
B
It's
an
enterprise,
it's
a
platform,
it's
awesome
and
the
datacenter
built
for
the
data
center.
But
you
know
if
you
want
to
run
any
kind
of
criminal
ease,
implementation,
you're,
building
you're,
doing
a
lot
of
there's
a
lot
of
operational
management
that
expects
you
know,
network
connectivity
be
consistent,
there's
a
lot
that
expects
you
to
run.
B
You
know
productize
and
turn
into
a
new
project,
and
so
we've
been
working
for
the
last
the
last
you
know
six
months
or
so
on
a
new
project
called
k3s,
which
is
a
certified
distro
of
kubernetes
and
all
that
Darin
take
it
from
here
and
really
introduce
you
to
create
k3s
Darren.
You
want
to
share
your
screen.
C
So
this
is
kind
of
the
result
of
a
lot
of
work
that
we'll
be
doing
for
a
while
and
I'll
kind
of
touch
on
some
of
those
like
where
some
of
this
stuff
came
from
from
different
projects
we
were
working
on,
but
so
introducing
k3s
right
up
front
I
wanted
to
kind
of
point
out.
So
it's
a
lightweight
certified
kubernetes
distro.
So
it's
like
you
know
technically
what
is
this
thing?
C
So
it's
really
just
a
smaller
distribution
of
kubernetes,
but
the
key
thing
that's
kind
of
different
is
because
there's
been
a
lot
of
projects
that
you
know,
I've
done
like
easier
to
install
kubernetes
or
smaller
footprint
or
whatever,
but
they're
pretty
much
all
oriented
towards
development.
So
it's
like
mini
cube
or
micro,
Cates
or
or
local
cube,
I
think
that
one's
think
on
basically
any
sit
in
them.
You
can
think
of
for
small.
C
The
project
probably
exists
like
kind
tiny,
pico
nano
whatever
there's
there's
been
a
lot
of
them,
but
the
key
thing
that's
different
about
KT
3s.
We
want
this
to
one
be
certified,
so
we've
already
gone
through
all
the
certification
and
we
pass
all
the
tests.
This
does
pass
the
conformance
and-
and
we
really
we
want
this
for
production,
so
this
is
not
going
to
be
just
you
know,
for
development
or
or
test.
This
is
really
for
you
know
real
production
workloads,
you
know,
hopefully
you
know,
as
we
get
more
and
more
users.
A
A
C
Supported
product,
so
you
know
how
big
is
it?
Well,
it's
the
binary.
Is
it's
just
a
single
binary
that
you
download
it's
40
megabytes
and
it
can
run
on
a
machine.
That's
got
512
megabytes
of
RAM
and
that's
kind
of
the
lowest
lowest
end.
I'll
show
all
demo
later
running
on
a
Raspberry
Pi
with
gigabyte
of
memory,
and
you
want
reducing
the
real
memory
usage
and
you
know
I
there.
C
But
it's
you
know
it's
significantly.
Smaller
I
mean
just
kind
of
anecdotally.
It's
about
1/4
the
size
of
a
regular
upstream
kubernetes
setup.
So
if
you,
if
you
run
like
cube
ATM,
just
set
up
a
single
node
cluster
with
cube
ATM
it
that
typically
takes
up
around
a
little
over
a
gigabyte
of
memory,
whereas
k3s
a
single
node
would
be
like
more
like
250.
C
So
the
binary
is
40
megabytes.
That
is
a
compressed
binary,
it's
gzip,
so
that
that
is
lying
a
little
bit.
It
has
it
it
won't
it
kind
of
dynamically
decompresses
as
you
as
you
run
it
so,
but
that
40
mega
by
binary
includes
not
just
kubernetes
but
also
everything
else
you
need
to
run
like.
So
you
can
have
you
know,
let's
say
like
an
alpine
system,
that's
something
that
doesn't
have
like
really
anything
on
in
the
OS,
and
you
can
just
run
the
k3s
binary.
C
You
don't
need
docker,
you
don't
know
any
hosts
utilities,
anything
it's
pretty
much.
We
just
need
the
kernel
and-
and
you
know
proper
mounts
for
like
praxis
and
dev.
So
you
just
run
this
one
binary.
It
includes
container
D,
flannel,
core
DNS,
C
and
I,
set
up
a
lot
of
the
additional
that
is
kind
of
outside
the
scope
of
core
kubernetes,
so
yeah.
C
So
we've
we've
packaged
this
differently,
like
one
of
the
reasons
why
a
regular
kubernetes
installation
takes
up
so
much
memory,
it's
just
like
the
duplication
of
the
processes
of
like
API
server,
controller
manager,
scheduler.
You
know
those
reasons
why
those
are
packaged
the
way
they
are
and
if
you're
running
a
full
scale
cloud
massive
cluster.
You
know
it
makes
sense
that
they're
separated
the
way
they
are.
But
when
you
get
down
to
smaller
clusters-
and
you
know
it
doesn't-
it
doesn't
actually
make
sense
of
anymore.
So
we've
combined
these
multiple
components
into
like
single
processes.
C
We're
still
keeping
proper
failure
boundaries.
You
know
we
didn't
reduce
the
reliability
or
kind
of
key
failure.
Scenarios
like
you
know,
decoupling
the
control
and
data
plane.
Those
are
all
staying
intact,
but
we've
we've
combined
certain
processes
to
reduce
them.
The
memory
footprint
we
added
sequel
Lite
in
addition
to
Etsy
D.
So
this
one
is.
This-
has
been
some
there's
a
project
that
I've
worked
on
for
a
long
time
with
kubernetes
of
experimenting
with
different
backends
to
kubernetes
at
CD
is
fine
I
mean
you
know
nothing
against
that
CD,
but
there's
there's
no
technical
reason.
C
C
So
that's
something
I've,
just
I've
experimented
with
a
lot
in
the
past
of
supporting
different
data
data
sources,
but
for
this
project
it
made
a
lot
of
sense
that
you
know,
since
we
already
had
kind
of
that
technology
and
figure
out
how
to
do
this.
It's
like
well
made
a
lot
of
sense
to
bring
sequel
like
this.
It's
significantly
smaller
memory
footprint
for
for
a
single,
you
know
a
single
master
cluster,
so
obviously
it's
equal
light
is
a
database
that
runs
in
memory.
C
So
there's
a
lot
of
reasons,
a
lot
of
use
cases
where
you
just
don't
care
that
much,
but
we
did
this
as
it's
just
an
option,
so
you
can
still
use
sed,
but
the
sequel
Lite
just
does
make
it
a
lot
smaller
and
it's
it's
quite
it's
quite
nice
to
deal
with
it
because
it's
everything
is
just
kind
of
stored
in
one
sequel
database.
You
can
use
sequel
my
tools
to
look
at
the
datastore,
it's
quite
easy
to
to
interact
with.
C
So
we
also
we've
released
this
like
we
want
to
make
sure
it
was
important
from
day
one
that
we
had
full
support
for
Intel
arm,
64
and
arm
v7,
so
you'll
find
that
there's
pretty
decent
support
for
arm
64
out
in
the
ecosystem,
but
arm
v7,
not
so
much
in
and
kind
of.
The
reason
for
that
is
when
you
look
at
like
running
in
kubernetes
on
a
Raspberry
Pi,
just
a
regular
or
like
upstream
one.
It's
really
it's
kind
of
too
fat.
It's
not
really
that
practical.
C
So
a
lot
of
people
don't
really
target
arm7,
and
so
that
wasn't
something.
Actually
we
struggled
a
lot
with,
as
we
were,
releasing
k3s
was
finding
like
projects
that
support
arm
v7,
making
sure
the
multi
arc
images
are
there.
You
know
working
on
things
like
that,
sixteen
various
bugs
and
like
container
D
around
arm
v7.
So
there's
quite
a
bit
of
effort,
I
mean
I,
would
say:
I
would
say
over
half
the
effort
of
like
trying
to
release
k3s
was
actually
getting
this
up
seamlessly
working
on
arm
so
from
day
one.
C
C
So
yeah.
Just
final
point
is
that
this
is
an
open-source
project.
It's
not
yet
a
product
with
Rancher
I'm
sure
you
know,
if
you
know,
there's
a
enough
demand
and
we're
already
seeing
a
good
amount
of
doing
we'll
turn
this
into
a
pro
a
product,
but
right
now
this
is
we're
just
announcing
the
open
source
project.
Also
just
kind
of
give
you
where
this
is
like.
This
is
not
what
we
would
consider.
Ga,
there's
still
things
that
we're
working
on
like
around
the
H
a
set
up
and
whatnot.
C
C
Beta
level,
you
know,
will
focus
on
GA.
You
know-
sometimes,
hopefully
this
year.
Okay,
so
you
know
so,
as
I
said,
you
know
what
what?
What
do
we
actually
do
to
make
this
happen
like
what
are
the
kind
of
technical
details
so
kind
of
on
the
left?
You
have
all
the
things
we
removed
on
the
right
is
what
we
added
so
fundamentally
to
make
things
smaller.
Like
less
memory,
we
had
to
remove
something
and
initially
the
way
this
project
first
started
was.
C
We
were
working
on
another
project
called
Rio,
which
is
this
experimental
project
that
I
started
last
year
that
focuses
on
the
server
l'espace
and
service
management
application,
and
when
we
were
doing
that,
we're
like
well.
We
want
this
to
be
so
easy
to
install
and,
let's
like,
let's
actually
embed
kubernetes
in
it,
so
you
don't
even
have
to
install
kubernetes.
So
what
we
did
was
we
just
took
kubernetes.
We
ripped
out
every
single
feature.
We
didn't
want
embedded
inside
of
Rio,
but
then
we
ended
up
realizing
we're
like
hey.
C
This
is
actually
really
neat,
because
Rio
ended
up
being
such
a
simple
way
to
get
the
kubernetes
stuff
running
that
you
know.
We
saw
a
lot
of
value
and
pulling
it
out
as
a
separate
project
beyond
the
fact
that,
like
we
were
seeing
demand
for
edge
and
all
those
other
things
is
a
you
know,
we
were
kind
of
happy
with
kind
of
technically
what
we
did,
but
at
the
time
that
we
did
it,
we
ripped
out
tons
of
features.
C
I
mean
things
that,
like
people
care
about
what
we
didn't
care
about
at
the
time,
so
we
kind
of
changed
the
scope.
We
said
well,
let's
figure
out
if
we
can
do
this
same,
like
reduction
in
size
but
but
break
the
core
functionality
of
the
cluster.
So
really
what
it
is
is
we
were
moved
legacy
functionality
like
some
of
the
really
old
API
versions,
because
the
you
know
the
way
Cooper
noise
does
things
like
the
API
is
will
go
through.
You
know
alpha
beta
and
stable,
so
some
of
the
older,
like
alpha,
api's,
we've
removed.
C
C
These
are
features
that
we
don't
we're
not
seen
very
often
and
as
as
Shannon
mentioned
in
the
beginning,
we
have
wrenches
got
something
20,000
clusters
that
that
people
are
running
on
so
we've
a
good
amount
of
users
that
we
even
seen
so
we
we
pulled
out
any
legacy
feature.
We
also
pulled
out
any
non
default
feature,
and
these
are
features
that
are
basically
not
even
available
in
cloud
managed,
kubernetes
instances
so
like
gke
and
eks,
and
those
ones
that
manage
the
the
the
control
plane
for
you.
C
They
don't
give
you
the
ability
to
change
like
admission
controllers
and
things
like
that,
and
so
these
are
features
that
you
know.
Typically,
you
don't
even
want
to
use
because
they're
not
very
portable,
but
there's
also
ways
to
do
these.
Now
that
don't
require
changing
the
configuration
of
a
map
server.
So,
like
emission
controllers,
we
pulled
out
a
lot
of
the
emission
controllers
that
aren't
enabled
by
default,
but
there's
still
the
webhook
value
like
that,
the
webhook
based
ones
and
so
there's
still
all
the
proper
extension
points.
C
So
you
can
still,
you
know,
bring
in
a
mission
controllers.
You
just
have
to
do
it
through
the
web
hook
approach,
which
is
the
preferred
approach,
anyways
anything
that
was
an
alpha
feature.
We
just
removed
they're,
not
enabled
by
default.
A
lot
of
them
aren't
even
stable.
That's
idea
of
alpha,
so
we
just
removed
them.
So
we
kind
of
the
intention
there
is
you
know
if
you
care
about
alpha
features,
you
know,
use
an
upstream
distribution
for
right
now.
Eventually,
if
those
Dafa
features
are
good,
they'll
become
beta
and
then
they
will
be
in
case.
C
Yes,
we
removed
in
cloud
provider
the
entry
cloud
provider,
so
we
didn't
get
rid
of
cloud
provider
like
in
general,
just
the
entry
one
so
like
the
the
the
approach
of
the
community
is
has
been.
You
know,
let's
pull
everything
out
of
the
built-in
binary
and
allow
them
to
be
deployed,
as
you
know,
separate
add-ons,
and
so
we've
done
that
with
so
anything.
That's
basically
entry
that
you
know
the
the
community
would
like
to
remove.
If
you
know,
basically,
we
just
went
ahead
and
removed
it.
C
So
a
couple
things
there
is
that
you
can
still
bring
in
a
cloud
provider,
but
the
kind
of
where
k3s
shines
is
not
going
to
be
super
deep
integration
with
clouds
like
you
might
as
well
just
use
gke
or
eks
or
digital
oceans.
You
know
whatever
the
cloud
provide
it
one,
those
are
going
to
give
me
the
best
experience.
C
K3S
is
more,
you
know
edge
or
you
know
all
these
other
use
cases,
so
those
things
just
didn't
really
make
sense.
So
also
we
remove
the
storage
drivers.
So
again
it's
the
same
thing
of
like
the
entry
storage
drivers.
They
can
all
be
replaced
now,
basically
by
CSI,
so
you
can
just
deploy
different
storage
drivers
through
CSI.
C
So
with
k3s
we
just
kind
of
say:
go
the
CSI
route,
there's
quite
a
few
of
them
available
already
and
then
the
last
thing
that
we
removed-
and
this
is
optional-
it
still
works,
but
by
default
we
don't
use
docker.
So
dr
adds
a
good
amount
of
memory.
Also,
you
know
it's
something
you
need
to
install
and
manage
on
the
host,
so
we
don't
by
default,
use
docker
we
instead
in
bed
and
use
container
d,
but
this
is
optional
because
I
know
there
is
still
a
lot
of
people
who
want
to
use
docker.
C
C
So
that's
how
we
got
the
memory
footprint
down
that
and
the
fact
that
we
combined
a
bunch
of
processes
and
so
the
things
that
we
added
to
k3s
to
make
it
you
know
easy
to
run
in
is
as
first
as
we
just
simplified
the
installation.
We
really
like
the
approach
of
like
most
deploying
kubernetes.
Is
you
run
some
type
of
installer,
so
I'm
like
something
which
goes
and
deploys
kubernetes
or
configures
it
or
whatever,
and
we
just
we
want
to
get
rid
of
that.
We
just
we
wanted
this
as
simple
as
possible.
C
Just
like
just
run
a
binary
like
you
just
set
up
a
system
the
unit
and
it
just
runs.
We
want
to
get
rid
of
all
that
that
kind
of
bootstrapping
and-
and
this
is
to
make
it
easier
to
kind
of
package
within
these
like
edge
scenarios
like
you
bake
it
into
the
image,
you
don't
need
a
lot
of
operations.
You
know
you
can
just
kind
of
you
know
just
package
it
up.
Really
easy
and
I'll
show
that
a
little
bit
more
when
I
demo,
so
we
had
its
equal
light.
C
So
I
talked
a
little
about
that
before
so
I
won't
do
that.
So
we
have
teal
TLS
management.
So
one
of
the
kind
of
the
hardest
parts
I
think
of
getting
kubernetes
setup
is
managing
all
the
certificates
and
everything,
and
so
by
default
we
just
generate
all
the
certificates
and
manage
it
already.
You
know
there's
already
you
could
work
going
on
even
in
the
community
of
simplifying
that
stuff,
like
that.
Cuba
can
do
like
certification
and
whatnot,
and
so
we
just
make
it
a
little
bit
easier.
C
C
We
do
automatic,
manifests
and
hard
shell,
shell
man,
if
it
management,
what
that
means
is
basically
I'll
show
this
later
is
that
we
just
make
it
really
easy
to
deploy
manifest
or
hound
charts,
and
this
is
makes
it
really
easy
for
packaging
and
different
use.
Cases
is
like
CI
and
I'll
show
that
when
I
demo
it-
and
so
the
last
thing
is
we
just-
and
we
include
all
the
other
dependencies
that
we
need
a
container
D
in
flannel
and
core
DNS
CN,
cn
I
stuff.
C
All
of
these
things,
everything's
optional,
anything
that
we
include
there's
some
way
to
disable
it
and
bring
your
own,
but
we
wanted
to
have
a
default
one
that
works.
You
know
kind
of
for
everyone,
so
you
can
like
disable,
flannel
and
install
a
different
CNI
driver
if
you
wanted
to,
but
for
the
most
part
you
know
a
lot
of
these
cases.
Flannel,
you
know
we're
twelve
okay,
so
this
is
just
a
picture
of
basically
how
how
it's
all
packaged
is
like.
We
just
basically
have
this.
C
We
present
kubernetes
is
just
like
an
architecture
of
a
server
and
an
agent.
So
if
you're
doing
a
simple
single
master,
you
just
have
one
server
and
so
that
it
just
has
one
process
that's
running
by
default.
When
you
run
the
server
like
when
you
run
the
binary,
it's
k3s
server,
it
will
automatically
launch
the
agent
on
the
same
machine.
So
it
becomes
a
single
node
cluster.
You
don't
have
to.
You
can
disable
it.
C
So
you
can
separate
the
server,
but
in
practice,
what
you
see
is
people
usually
run
kubernetes
on
the
management
plane
too,
so
they
can
manage
control,
plane,
components
and
then,
additionally,
for
like
the
edge
use
case,
you
want
to
reduce
the
number
of
nodes
you
have
anyway.
So
why
would
you
dedicate
you
know
a
single
full
machine
just
to
the
server?
C
C
This
tunnel
are
the
the
to
proxy,
the
cubelet
flannel,
and
then
it's
going
to
launch
container
D
and
then
there's
this
component
in
there
that
the
tunnel
proxy
and
the
reason
what
this
is
is
like
is
the
is
the
way
kubernetes
works
today
or
like
upstream
kubernetes
works
is
there's
bi-directional
communication
from
the
control
plane
to
the
cubelet.
This
means
that
there's
both
outbound
and
inbound
TCP
connections
going
between
those.
So
the
inbound
TCP
connections
to
the
nodes
for
a
lot
of
scenarios
causes
problems.
C
It's
a
little
harder
to
manage
from
a
security
and
networking
perspective.
So
what
this
tunnel
proxy
is
is
it
sets
up
a
tunnel
from
the
agents
to
the
server,
and
it
makes
makes
it
so
that
all
communication
from
the
nodes
is
outbound,
and
so
this
is
something
we've
we've
added
in
is
just
as
part
of
k3s.
So
this
makes
the
networking
you
know
kind
of
setting
this
up
in
a
secure
manner,
a
lot
easier,
because
the
nodes
just
need
an
album
connection
and
there's
no
other
listening
ports
like
byte
by
default.
C
This
is
supposed
to
be
secure.
The
only
port
that
you
really
have
to
worry
about
is
the
VX
LAN
port
and
like
we're
really
trying
to
see
if
we
can
switch
to
like
IPSec
or
wire
guard,
to
make
the
networking
so
by
default.
But
right
now
there
is
from
one
security
perspective.
Is
you
have
to
worry
about
that?
The
X
LAN
port,
it's
like
might
be
four
seven.
C
B
C
I
mean
so
that
so
we
didn't,
we
haven't
like
fundamentally
change.
What
kookaburras
is
so
you
can
still
have
storage
drivers.
You
just
need
to
pull
them
in
through
CSI
and
we
actually
have
a
really
nice
local
storage
driver
that
we've
developed-
it's
not
included
in
k3s,
but
it's
really
easy
to
develop
or
to
deploy.
That
gives
you
a
local
storage
version
of
PBS.
C
You
just
need
to
have
some
some
driver,
and
so
we
don't
right
now
include
a
storage
driver.
That
is
something
I've
been
toying
with.
The
idea
of
you
know
whether
or
not
we
should
include
like
a
local
storage
one,
so
you
can
least
use
PBS
out
of
the
box,
but
if
you're
looking
at
persistence,
you
know
you
really
need
to
give
some
consideration.
So
you
know
what
type
of
storage
but
so
persistence
works.
You
know
you
just
have
to
give
it
that
extra
thought
that
is
required
in
general
with
kubernetes.
B
Think
that's
a
good
I
mean
just
that
applies
to
almost
everything
as
well.
There's
other
questions
about
you
know
the
the
entry
cloud
providers
again
all
these
things
can
be
just
added
back
in.
If
you
need
them,
we
just
removed
them
by
default,
because
otherwise
they
take
up
a
lot
of
space,
their
bundle.
Then
they
just
kind
of
fatten
up
what
gets
deployed,
because
you
include
all
the
cloud
providers
instead
of
the
ones
that
you,
if
you
do,
want
to
run
this
or
thought
you
just
Google
or
Amazon
digitalocean.
C
C
No,
that's
that
not
the
best
support
for
that
right
now.
It's
so
it's
all
kind
of
internal
management.
But
it's
not
it's
not
as
it's
not
as
like
kind
of
baked
in
as
swarm
or
there's
like
you
know,
no
chance
you
can
kind
of
get
to
it,
because
we
are
still
launching
like
the
regular
components
with
the
regular
command-line
arguments
and
whatnot,
so
we're
just
kind
of
generating
files
and
putting
them
on
disk,
and
so
it's
possible.
If
you
don't
want
to
use
the
generated
ones,
you
can
just
put
in
your
own
certificates.
C
There's
not
like
a
kind
of
like
I
guess
you
could
probably
get
clever
with
cert
manager,
the
cert
manager
you
can
allow.
You
could
probably
wire
that
up
to
do
external
CA
integration,
so
I
mean
there's
no
there's
a
lot
you
can
do
and
in
the
end,
we're
not
it's
not
so
I!
Think
kind
of
the
short
answer
there
is
it's
like
it's
not
heavily
baked
in.
You
can
get
to
it.
C
If
you
want
to
do
more
advanced
use
cases
and
we're
definitely
open,
you
know
if
you're
there's
something
you
want
to
see
supported
you
put
in
an
issue.
This
is
all
about.
I
mean
there's
all
these
projects
that
we
launched
with
the
Rancher.
This
all
all
about
the
users.
We
just
you
know
whatever
it
is.
You
want
to
do,
let's
you
know
figure
out
how
to
support
it.
C
C
You
know
that's
kind
of
like
the
most
obvious
thing
we
can
kind
of
market
and
put
this
towards,
but
like
k3s
I
think
is
so
much
there's
a
lot
more
potential.
What
you
can
do
you
just
have
this?
You
know
really
simple
easy
to
run
kubernetes
and
so
that
we
really
think
there's
a
lot
more
use
cases
for
it
in
so
we're
really.
You
know
we're
kind
of
excited
to
see
what
people
do
with
it.
You
know
there's
things
that
we're
looking
at
because,
like
you
know,
we
realize
is
like
hey.
C
This
actually
ends
up
being
a
really
nice
replacement
mini
kit.
If
we
were
to
put
a
little
bit
of
UX
around
it
or
whatnot,
because
it's
smaller
it
uses
less
CPU.
You
know
it's
nicer
on
your
laptop,
so
there's
different
things.
So,
like
CI
environments,
we
started
heavily
using
that
internally,
so
anytime,
like
more
running
CI
and
we
need
to
launch
kubernetes
and
run
so
test
against,
and
we
swap
that
out
for
kts.
So
there's
a
lot
of
different
use
cases,
it's
not
just
for
edge
and
arm.
You
know,
there's
a
lot
of
things.
C
You
could
kind
of
figure
out
where
you
might
want
to
put
k3s
okay,
so
I'm
gonna
go
through
the
demo.
This.
This
is
on
the
website.
This
information
here,
but
I'm
gonna
go
through
and
kind
of
demo
show
you
a
little
bit
of
this
in
action
and
that
demo
is
actually
very
long
because
I
mean
honestly
there's
not
there's
not
a
lot
to
this
in
terms
of
a
it's
just
getting
the
kubernetes
cluster
up
and
ready.
So
I
just
want
to
first
show
like
this
is
a
raspberry
pi.
C
C
You
know
so
like
here's
like
piglet,
you
know,
so
this
is
all
running
on
a
Raspberry
Pi.
So
you
know
I
got
this.
This
full
thing
set
up,
so
you
can
see
from
like
the
kind
of
free
memory.
A
Raspberry
Pi
has
a
gig
of
memory,
so
I
still
have
free
half
half
of
a
gig,
so
I
have
kubernetes,
plus
open
faz,
and
then
the
couple
functions
deployed
and
I
still
got
a
half
half
a
gig
free.
So
it's
you
know
it's
not
like.
C
We
made
kubernetes
super
tiny
I
mean
it's
still
a
couple
hundred
megabytes,
but
even
on
something
like
a
Raspberry
Pi,
you
still
have
a
good
amount
of
space
for
your
applications
to
manage
it.
So
it
is
even
feasible
to
run
it
on
a
device.
This
small
and
we
do
test
it
on
them.
We
have
these.
My
favorite
arm
device
I
have
is
a
beagle
beagle
bone
pocket.
It's
a
little
v,
512
megabyte
arm
v7
device,
and
it
will
run
on
those
two,
but
it
ends
up
taking
up
the
majority
of
the
memory.
C
C
Command
line
said
that
one
yeah,
you
have
to
add
these
two
parameters.
If
you
don't
add
these
two
parameters
on
a
Raspberry,
Pi
k3s
will
actually
tell
you
it'll
fail
to
start
and
I'll
say
hey,
you
need
to
add
these
parameters,
but
I
just
want
to
quickly
show
that
I'm
not
going
to
demo
a
lot
on
the
Raspberry
Pi
because
things
just
move
slower,
but
so
it
is
actually
feasible
to
run
it.
There's
that
small
of
a
device
they're
still
memory
free.
Let
me
see
if
I
look
at
the
you
know.
C
It's
not
like
it's
sitting
here.
The
load
average
load
average
is
still
below
one.
It's
not
like
it's
killing,
killing
the
machine,
but
we
are
working
on
trying
to
figure
out
how
to
get
this
even
as
less
CPU.
You
know
figuring
out.
This
is
just
kind
of
kubernetes,
regular
churn
and
so
we're
doing
more
work
on
trying
to
get
the
CPU
even
lower
okay,
so
I'm
gonna
walk
through
the
installation
on
these
are
arm
64
boxes,
so
these
are
a
little
bigger
and
faster.
These
are
running
in
Amazon.
C
We
do
a
lot
of
our
arm
stuff
on
Amazon
or
packet.
Both
of
those
are
great
providers
to
get
arm
servers,
so
the
simplest
way
to
get
up
and
running
is,
is
the
curl
script
and
I'll
show
that,
but
then
that
kind
of
hides
what's
going
on
so
I'll,
do
it
manually
too,
but
just
to
show
you
you
know
what's
going
on
here.
So
if
you
run
this
curl
script,
this
basically
sets
up
an
uninstall
script
which
is
kind
of
nice,
so
you
can
get
rid
of
it
easily,
but
basically
we
just
copy
the
binary.
C
We
sim
link
these
two
programs
to
the
binary,
so
you
get
cube
CTO
and
CRI
CTO
CRI
CTL,
CRI
CTL
is
kind
of
like
your
docker
PS.
It's
the
way
you
can
look
at
containers
locally
on
the
node,
then
we
register
it
with
system
D
and
then
start
up
system
D,
and
so
in
theory
this
should
work.
Maybe
it
takes
about
yeah,
okay.
So
what
re
it's
up
its
up
and
running?
So
so
let
me
explain
to
you
like
by
default,
what
you
get
in
the
instant
loop,
what
you
kind
of
get
by
default?
C
What's
deployed
so
in
a
regular
cluster,
what
we
do
is
redeploy,
so
we
have
courtney
NS
running
and
then
this
is
a
so
we
have
coordinating
s
running
for
you
know
the
dns
support
we
install
traffic.
So
initially
we
started
with
the
engine
x,
ingress
controller,
but
it's
not
well.
It's
not
supported
at
all
on
arm
v7,
so
traffic
is
supported
on
arm
v7,
plus
it's
actually
a
lighter
weight,
less
memory
than
the
engine
X
controller,
so
we
right
now
default
to
traffic.
Again.
Everything
in
here
is
optional.
C
You
can
flip
them
off
and
bring
your
own
thing.
If
you
don't
like
this,
so
we
automatically
deployed
traffic.
So
you
get
an
ingress
controller,
already
working
I
and
then
you'll
notice
that
this
has
the
external
IP
already
set
up.
So
what
we
have
running
is
we
have
a
very
simple
service
load,
balancer,
that
it
works
based
off
of
host
ports
so
because
that's
always
been
a
problem
with,
like
you,
don't
have
cloud
integration,
how
do
you
the
service
load,
balancer
and
congreso?
C
We
just
developed
a
very
simple
one
that
like
well
just
if
there's
a
host
port
available,
it'll
pick
one
of
the
nodes
and
use
a
host
port,
and
so
this
works
really
well
for
these
smaller
sets,
you
know
kind
of
smaller
clusters,
but
again
you
can.
You
can
bring
in
a
different
service
load
balancer
like
metal,
lb,
or
something
like
that.
C
So
those
are.
The
two
main
things
are
running
is
traffic
and
Courtney
ns,
and
then
this
this
is
the
load
balancer.
That's
that's
doing
the
host
port
thing.
This
basically
takes
up
no
memory,
it's
just
it's
just
a
it's
like
a
pause
container,
because
it's
really
just
done
with
IP
tables
all
right,
and
then
this
is
the
helm,
install
okay,
so
the
helm,
install
so
okay,
I'll,
come
back
to
the
helm,
install
once
I
show
that
the
installation
so
I'm
gonna
uninstall,
this
I'm
gonna
just
delete
it
just
so
you
just
run
that
script.
C
Trying
to
get
that
that
working
a
little
better,
let
me
kill
those
ones
so
I
know
everything's
totally
gone
oops,
okay,
but
I'm
going
to
show
you
kind
of
the
manual
setup
just
so
you
see
because
the
script
to
kind
of
feel
like
it
makes
it
super
simple,
but
you
know
kind
of
office
gates
what's
going
on,
and
so,
if
you're
doing
this
manually,
you
just
go
to
our
releases.
Go
to
our
latest
release.
Just
look
for
the
k3s
binary!
C
You
want
so
I'm
gonna
do
arm
arm
64,
and
so
you
can
see
here
this
one's
it's
32
megabytes
is
actually
less
than
40.
So
we're
just
going
to
download
this
move
that
over
to
k3,
yes,
make
it
executable,
and
then
we
just
run
it.
So
we
just
take
a
3s
server,
so
this
isn't
a
launch
the
server
and
then
additionally,
it's
going
to
register
itself
as
as
a
local
node.
So,
okay,
so
the
server
started.
Now
it's
launching
the
agent,
and
so
the
agent
will
take
a
little
bit
like
get
the
cedar
and
whatnot.
C
But
when
this
launches
it
runs
this
this
command.
It
outputs
this
this
information
here.
So
it's
says
if
you
want
to
join
a
node
to
the
cluster,
just
run
this
command
and
you
need
to
give
it
the
node
token-
and
the
note
token
is
here
in
this
file,
so
if
I
cat-
that
it
gives
you
this
this
big
long
token
here.
So
it's
an
approach
very
similar,
the
cube
ATM,
but
it's
slightly
different
because
it
has
to
it
does
some
magic
to
set
up
the
TLS
certs
and
do
things
securely.
C
B
A
C
C
Anyways
so
because
I'm
kind
of
running
out
of
time
here,
but
so
the
it's
it's
not
it's,
not
it's
not
actually,
because
I
think
I
didn't
set
up
the
security
gears
between
these
two
machines,
but
so,
basically
that
would
just
join
it.
Does
the
exact
same
thing
like
you
would
see
kind
of
this
output
here
where
it
connects
and
it
joins
together.
C
If
you
want
to
use
K
III
assets
in
a
CI
environment,
you
can
just
basically
create
a
directory
put
some
yamo
files
in
there
run
the
binary
it
launches
the
cluster
automatically
deploys,
and
then
you
can
start
running
tests
and
whatnot,
and
so
you
can
see
here
like
that's
how
we
deploy
core
DNS.
We
were
running
core
DNS,
but
a
lot
of
things
are
practiced
as
helm
charts.
C
So
we
have
automatic
integration
with
helm,
charts
too,
or
you
can
just
go
and
create
a
manifest
here
for
a
helmet
art
you
just
put
like
what
the
chart
is.
If
there's
a
repo,
that's
you
know
this
one
is
in
the
upstream
repo,
but
you
can
put
repo
whatever
parameters
you
want
to
set
and
then
they'll
deploy
the
helmet
chart
for
you.
So
the
way
we
actually
deploy
traffic
is
through
a
helm
home
chart,
so
we
have
like
integrated.
C
This
is
just
to
make
it
easier
to
package
and
get
you
know
kind
of
software
bootstrapped
and
run
on
the
cluster,
so
I
see.
So
that's
that's
most
of
so.
The
other
thing
I
just
wanted
to
point
out
was
that
for
the
installation,
the
way
we
did
this
is
we.
We
ran
the
agent
fortune
kind
of
stinks.
This
failed
but
such
as
live
demos
yeah.
So
you
can
see
the
connection
time
now
because
there's
it's
either
a
networking
issue
or
an
obscure
group
setup.
C
But
I
wanted
to
point
out
that
the
way
this
is
done
to
bring
up
a
cluster
is
you
run
the
server
it
generates.
This
token,
then
you
use
touken
to
join
the
cluster.
That
makes
it
difficult
to
easily
automate
bootstrapping
fold
cluster.
So
there
is
another
approach.
Instead
of
using
the
generated
cluster,
you
can
pass
a
cluster
secret,
a
shared
secret,
which
is
just
any
random
string.
C
You
want
to
come
up
with,
so
if
you
give
the
same
string
to
the
server
and
the
nodes,
then
you
don't
have
to
like
do
that
two-step
process,
so
it
makes
it
super
easy
to
bring
up
a
cluster
with
a
shared
secret,
so
you
can
start
getting
in,
like
you
can
start
doing
some
really
interesting
things
with
this
like,
for
example,
if
you're
you
know
into
like
docker
and
docker.
This
is
like
a
composed
file
to
launch
a
a
3s
cluster
use
the
server
and
like
here's
the
nodes.
C
C
B
A
So
what
I
just
want
to
tell
the
audience
is
that
we're
gonna
have
the
slides
and
recording
of
this
posted
later
today.
So
you
know,
look
for
those!
We
also
look
forward
to
seeing
you
at
our
next
scenes.
You
have
webinar,
so
you
everybody
have
a
great
day.
I'm
gonna,
let
it
Shannon
take
over
on
the
Q&A
I
see
that
we
have
11
questions
right
now.
Shannon
all
really
good
questions
too.
So
I
will
let
you
answer.
Those
I
have
to
go
to
another
meeting
thanks.
B
C
Yeah
I
mean
I
think
the
short
of
it
is
no,
the
the
it's.
It's
simple
enough
to
just
run
a
bunch
of
server
containers,
so
you
can
have
like
one.
You
know
one
server
machine
and
then
just
launch
a
bunch
of
containers
which
are
like
the
server
processes
and
then
join
those
to
those.
But
we
don't.
We
don't
kind
of
plan
on
I'm,
doing
multi,
tenant,
tenants,
I
guess
so.
B
There
was
another
question
at
DW
asked:
can
you
say
something
about
how
the
agent
compares
to
virtual
cubelet?
Could
the
agent
replace
the
virtual
cubelet
of
the
future
for
the
hybrid
environment,
VK
seems
to
struggle
with
the
two-way
communication,
as
you
mentioned,
for
example,
also,
could
an
agent
join
a
conventional
cluster
thanks
Doug.
C
The
agent
can't
join
a
conventional
cluster
because
the
agent
relies
on
the
tunneling
and
the
bootstrap
process
that
we
have
built
into
k3s.
So
the
kind
of
the
wrapper
around
the
agent
won't
work.
But,
having
said
that
is
we
do
besides.
So
there's
two
things
about
about
k3s
is
we
made
it
smaller
and
then
we
wrapped
it
in
this
installer.
C
So
that's
a
very
advanced
use
case,
but
that
that's
basically
this
gives
you
just
a
smaller
kubernetes
like
the
hypercube
of
kubernetes
is
I,
think
around
220
megabytes
and
ours
is
70,
so
you
can
just
see
like
ours
is
much
smaller.
So,
okay,
then
the
last
thing
is
a
virtual
cubelet.
The
virtual
cubelet
is
kind
of
a
different
use
case,
so
a
virtual
cube
will
work
with
k3s,
but
the
virtual
cubelet
is
it
really
is
a
it's
for
kind
of
virtualizing.
C
The
idea
of
a
note
you
know
so,
like
an
ode
becomes,
you
know,
an
ode
is
basically
like
a
huge
cluster
or
a
resource
pool.
So
that's
kind
of
a
different
use
case,
and
so
there's
there's
not
exactly
an
overlap,
but
the
virtual
cube
will
still
work,
we'll
work
with
k3s
and
in
different
work.
We're
actually
doing
a
lot
of
interesting
things
around
virtual
couplets
in
different
kind
of
probe
research
stuff
that
we're
doing
right
now.
Awesome.
B
C
Finals
for
the
overlay
networking
and
so
from
the
north-south.
It's
a
combination
of
the
ingress
controller
and
the
service
load
balancer,
so
the
service
load
balancer,
the
one
that's
built
in
is
just
based
off
of
IEP
tables,
and
you
know,
and
then
the
ingress
Controller
is
based
off
is
that's
traffic.
So
that's
going
to
give
you
the
north-south
traffic
or
no
South.
B
C
We
don't
include
Etsy
D
in
the
package
right
now.
We
might
because
we're
still
actually
working
on
the
H,
a
solution
because
we're
we're
gonna
make
the
H
a
kind
of
very
simple.
Just
like
we've
done
the
single
master,
it's
very
simple!
So
right
now
the
full
EDD
is
not
included
in
the
package
like.
So,
if
you
want
to
use
sed,
you
would
have
to
basically
configure
it
yourself
right
now
out
just
due
to
a
little
bug.
It
doesn't
actually
technically
work,
but
it
will
be
fully
supported
more
very
quickly,
we'll
fix
that.
C
But
right
now
the
sed
management.
If
you
were
to
do
it,
you
would
be
spitting
up
and
managing
sed
yourself.
So
that's
what
I'm
saying
is
like
we
want
to
make
that
even
easier,
so
we
will
probably
include
at
CD
into
k3s
we're
not
sure
we
don't
think
it
will
add
a
huge
amount
to
the
binary
size
and
then
in
the
memory
size.
The
way
we
plan
on
running
it,
because
we've
done
this
in
other
projects.
B
C
Ya,
I'm
not
super
suits,
so
those
are
kind
of
different
at
taking
a
completely
different
approach
and
I.
Think
that's
not
saying,
like
the
virtual
tubular
is
kind
of
a
whole
different
approach
on
how
to
do,
and
that's
kind
of
been
like
a
lot
of
people's
the
thing
of
like
well.
How
can
I
do
IOT,
but
what
that
does
is
like.
Basically,
the
virtual
cubelet
is
like.
C
But,
as
I
said
before,
we
are
doing
some
interesting
things
with
virtual
cubelet,
but
I
think
that's
just
a
little
bit
farther
out
there
in
the
danger,
with
the
virtual
cubelet
right
now.
Is
the
fact
that
it
is
like
this
big
proprietary
wedge
and
so
we're
trying
to
figure
out?
You
know
kind
of
what
to
do
around
that.
C
Is
a
great
question:
I
have
no
clue
so
I
am
a
big
Chrome
OS
user
and
there
is
a
cgroups
issue
in
general.
There's
one
of
the
things
the
CPU
set
C
group.
That
makes
it
so.
You
can't
typically
run
like
Oh
containers
on
on
the
device,
but
I've
been
looking
to
work
around
that
it's
like
it's
just
a
patch
and
run
see
so
I
would
say
it's
probably
feasible
to
do
it,
but
it
won't
work
right
now,
but
I'd
be
very,
very
interested
in
it
because
I've
been
trying
so
they're.
C
The
reason
why
I
mention
Chrome
OS
is
Chrome,
OS
and
Android
are
very
close
to
each
other.
Now,
the
the
way
a
Chrome
OS
kernel
is
set
up
is
very
similar
to
the
two
android
kernels,
but
I
think
it
might
actually
be
possible.
I'll
be
interested
interested
if
to
try
that
now,
when
I
probably
try
it
now,
but
I
know
won't
work
at
the
moment,
but
I
think
it's.
This
feasible.
B
C
C
So
if
you
have
docker
for
Mac,
then
you
can
run
this
because
we
were
looking
at
this,
and
this
was
not
our
initial
focus
like
the
laptop
or
the
development
environment,
but
we
actually
realized
it's
like
well.
K3S
on
the
laptop
could
be
really
nice,
so
I've
been
toying
around
with
the
idea
of
creating
a
kind
of
separate
wrapper
around
k3s
or
designed
just
for
laptops.
C
C
Yeah
I
can't
I
can't
really
commit
to
that,
so
that
doll
depends
on
user
demand.
I
mean
to
be
honest.
What
we've
seen
like,
as
you
know,
we're
just
actually
officially
announcing
this
project
today,
but
people
have
stumbled
upon
it
as
we've
been
developing
it
because
we
do
everything
in
the
open.
So
even
though
we've
never
really
announced
it,
it's
been
out
there
and
github,
and
we've
had
quite
a
bit
of
interest
in
it
without
marketing
it
or
anything.
So
it's
all
based
on
user
demand.
C
B
You
need
something
you
seem
like
an
immediate
use
case.
We
have
like
there's
like
30
people
that
are
doing
interesting
tests
on
this
stuff
right
now,
all
of
which
are
asking
more
GA,
Suen
or
so.
But
you
know
the
more
people
need
it.
The
more
we
can
put
resources
on
getting
it
brought
forward,
Henry
reassess.
How
do
you
operate
under
the
intermittent
in
connection
for
certain
nodes.
C
Right,
yeah
so
that
one
I'd
say
like
we
haven't
we're
still
working
on
it.
That's
kind
of
I
would
say
on
the
list
of
like
to
do
for
GA.
So
right
now
you
just
basically
we
have
to
tweak
some
of
the
intervals,
so
it
will
I
mean
it
it.
It
does
work.
The
problem
you
have
to
just
be
afraid
of
is
is
if
a
node
goes
unavailable
for
a
certain
amount
of
time,
kubernetes
to
start
moving
around
workloads
and
so
we're
trying
to
figure
out.
C
B
There
was
a
another
question
from
Hussain
who
kind
of
followed
up
on
his
earlier
question
and
he
was
saying
well
if
a
starts
with
sequel
light
in
a
single
node,
a
net
CD
then
using
the
cluster
node.
Could
you
start
with?
How
do
you
start
with
single
note
and
then
scale
that
out
to
say
three
or
five
nodes
in
the
future?
So,
but
if
you
start
with
yeah.
C
C
So
I
mean
the
sequel
light.
Is
it's
a
very
simple
kind
of
data
structure?
That's
in
there,
because
it's
just
key
value
so
so
as
they're
saying
we're
still
technically
working
on
the
AJ
solution.
So
it's
not
really
like
working
right
now.
So
we
are
as
we
implement
that
we
will.
We
will
provide
a
means
to
basically
move
from
the
sequel
line
to
that
CD,
which
is
just
basically
a
script,
to
basically
export
the
database
and
put
it
into
sequel,
Lite
or
into
ed.
C
Cd
is
a
very,
very
simple
structure,
so
it's
not
very
hard
to
move
between
the
two,
the
kind
of
the
side
effective
and
what
I
was
saying
before
is
that
the
the
behavior
of
Etsy
D
is
so
baked
into
kubernetes
that
to
implement
it
on
top
of
a
relational
database,
you
actually
have
to
replicate
the
log
structure
of
a
bet
CD,
so
it
actually
mimics
the
structure
of
that
CD
pretty
closely.
So
it's
pretty
easy
to
move
between
the
two
with
just
a
migration
script,
but
that
doesn't
exist
today
as
well.
B
Awesome
next
question
was
from
Aaron.
Oh,
he
said,
and
by
the
way,
there's
more
questions
coming
in
as
I
asked.
So
that's
why
I
know
we
said
they
had
11
before
we
somehow
still
have
11,
so
we're
just
not
feel
free
to
keep
firing
him
in
just
FYI
keep
going
Arno
said
it's
computer
D
image
download
process
more
edge
efficient
than
the
dr.
parallel
mode.
Oh.
C
C
C
C
C
The
senses
situations
were
you
know
because
of
like
having
to
use
like
IP
and
IP
or
or
or
you
know,
BGP
or
whatnot
like
it
requires
some
integration
with
the
with
the
network
setup
forward,
I'm
just
like
so
I
and
maybe
I
could
be
wrong
in
this
area,
but
I
don't
think
it's
just
basically
kind
of
generically
will
work
always
everywhere,
but
it
is
a
great
great
solution.
So
flannel,
the
final
VX
I'm
one
is
kind
of
the
lowest
common
denominator
denominator.
C
It
works
everywhere,
but
one
of
the
things
is,
we
don't
have
right
now,
a
network
policy
implementation
out
of
the
box,
and
so
that's
something
that
we
would
look
at
pulling
from
calico,
which
is
I,
think,
is
technically
Felix,
which
is
the
the
policy
manager
so
that
we
would
be
looking.
But
but
so
it's
just
the
main
reason
we're
pulling.
Calico
is
just
the
the
I
just
say:
I,
don't
think
it's
I,
don't
think
it
just
works
everywhere,
just
kind
of
out
of
the
box.
C
C
So
not
out
of
the
box,
so
you
are
gonna,
have
to
deploy
a
policy.
Some
component
that
enforces
the
policy,
so
I
think
that's
something
reasonable
that
we
would.
We
could
include,
indicate
three
s
or
make
it
as
like
an
option
to
turn
on
a
lot
of
people.
Don't
use
network
policy.
So
it's
it's
a
little
harder
for,
like
the
edge
use
case.
B
C
So
right
now
I
mean
it's,
it's
really
it's
just
as
binary.
So
if
you
were,
if
you
replace
the
binary
and
restart
so
on
the
server,
just
replace
the
binary
restart
on
the
agents,
replace
it
and
restart.
So
that's
all
it
is
right
now.
So
the
idea
is
that
that's
pretty
simple,
it's
just
a
binary
and
then
we
can
kind
of
wrap
that.
However,
we
want,
like
you,
know,
put
it
into
a
proper
OS
packages
or
build
a
auto
upgrade
process,
some
type
of
thing
scholar
or
something
so
right.
C
Now
it's
a
pretty
basic
unit
of
just
it's
a
binary,
so
just
replace
it
and
restart,
and
you
should
be
good
to
go
the
the
normal
like
if
I
kill.
If
I
have
all
the
containers
running
and
I
kill
the
agent,
it
won't
kill
the
containers.
They'll
still
keep
running,
so
you
just
put
the
new
one
in
and
then
then
it
runs
anywhere.
Okay,.
B
I
think
we're
certainly
past
the
time,
but
we've
got
four
or
five
more
questions
left
we'll
try
to
get
through
these
and
and
then
wrap
up
pretty
quickly.
Stephen
said:
are
there
any
OS
dependencies
on
the
nodes?
Could
I
use
something
like
tiny
core
Linux?
Would
it
still
work
with
docker
be
the
flag
even
on
tiny
core,
so.
C
There's
really
no
OS
dependencies,
so
I'll
show
you
just
a
little
bit
kind
of
under
the
hood.
How
this
works
is
when
we
kts
is
actually
an
archive
that
extracts
and
runs
out
of
that,
and
so
it
goes
into
this
data
directory
and
it's
hashed
according
to
the
version,
and
so
basically
this
is
everything
we
have.
It's
pretty
much
a
full
busy
box
set
up
with
like
key
things
like
s,
okat
and
run
C
picks.
So
basically
we
have
pretty
much
like
a
full
micro
distribution
of
Linux,
so
there's
no
OS
dependencies
at
all.
C
So
the
key
things,
though,
is
that
like
is
like
the
cysts
like
sis
seat
like
these
seat
groups,
need
to
be
set
up
properly,
the
the
so
the
Praxis
and
dev
mounts
need
to
be
set
up
properly.
So
most
Linux
distributions
do
this
already
something
like
tiny
core,
probably
won't,
but
if
you,
you
know,
know
how
to
mount
these,
and
you
can
do
that.
But
so
that's,
basically
the
only
dependency.
So
the
binary
itself
is
even
like
I'm
statically
compiled
against
muscle.
So
we
don't.
B
Jethro
kind
of
has
to
follow
it.
Sort
of
the
same
question
slight
varies
is
what
do
I
need
to
start
with
k3s
when
I
have
an
OS,
say
bun.
The
only
thing
I
need
to
do
download
this
and
run
it
like
you
demoed
or
do
I
need
some
other
settings
to
do
for
first
on
Raspberry
Pi,
you
should
something
about
boot.
Settings
I
can't
find
anything
in
the
readme
on
github
about
this
yeah.
C
So,
there's
really
nothing.
You
need
to
do
for
most
modern
Linux
distributions
so
like
if
you're
using
a
mainstream
normal
one,
you
know
Ubuntu
Red
Hat,
you
know
sooo
say
whatever
that
all
the
normal
kind
of
big
ones
or
whatever
debian
it
will
just
work
you,
so
you
should
just
be
able
to
download
it
and
run
it
just
run
that
curl
script
or
manually
put
the
binary
in
place.
C
So
Raspberry
Pi
is
a
little
unique
because
they've
flipped
off
some
of
the
default
settings
for
memory
reasons,
we'll
probably
will
update
the
docs
to
put
some
of
those
steps
for
the
Raspberry
Pi.
It's
really
just
that
one
setting,
but
that
the
binary
will
warn
you
anyways.
So
the
installation
is
as
much
as
possible.
We
want
to
make
it
so
it's
really
just
put
that
binary
and
run
in
so
so
you
know
it
includes
everything
that
then
it
needs
awesome.
C
So
I
mean
so
this
is
the
kind
of
normal
kubernetes
architecture.
Stuff
is
like
the
kubernetes
architecture
is
well
done
in
that.
Could
the
control
plane
and
the
data
plane
are
separated
such
that
like,
if
you
take
down
any
of
like,
let's
say
the
the
server
or
you
take
down
the
agent,
those
are
considered
considered
like
control,
plane,
components
all
of
your
workload,
the
networking
the
load
balancer
all
that
stuff
will
continue
to
work,
so
your
applications
are
still
up
and
running,
even
if
you
kill
the
the
server
or
the
agents.
C
If
those
are
down,
you
obviously
can't
interact
with
kubernetes
and
do
anything,
but
you
know
that's
kind
of
expected.
If
I
bring
down
the
server
and
replace
it
and
upgrade
it,
it
will
work
just
fine,
so
you
should
be
able
to
upgrade
this
thing
with
no
impact
to
your
applications.
That's
the
general
idea,
sometimes,
depending
on
how
big
of
an
upgrade
it
is
between
the
internal
kubernetes
versions.
Pods
might
get
reloaded
so
that
kind
of
depends
on
the
kubernetes
if
they
make
a
change
that
they
have
to
reload
pods.
C
B
B
Who
wants
to
kind
of
find
this
stuff
out
there?
You
know
we've
got
a
website
set
up
for
k3
SEO.
You
can
also
find
it
on
github.
Just
github.com,
slash,
rancher,
slash
k3s,
and
you
know,
as
I
mentioned,
Darrin
and
I
are
more
than
happy
to
take
your
questions
whenever
you
have
them
just
shoot
us
a
note
on
slack
or
zoom
or
wherever
it's
convenient
for
you
with
that
I
think
we're
wrapped
up
Kaitlyn.
Thank
you
so
much
for
inviting
us
to
present
and
interpro
seeing
us
today.